@ North-Holland Publishing Company

Chemical Physics 48 (1980) 157-173

1.

A COMPLETE ACTIVE SPACES& METHOD- (CA&F) FORMULATED SUPER-CT APPROACH Bjiirn 0. ROOS, Peter R. TAYLOR?
Department of Physicul Chemistry

U&G

A DEiiiITY

MATRM.

: :

2, Chemica! Centre, S-220 07 Lund 7, SGeden

and Per E.M. SIEGBAHN

Institute of TheoreticalPhysics, S-113 46 Stockholm, Sweden

Received 19 November 1979

A density matrix formulation of the super-C1 MCSCF method is prmetited.The MC expansion is assumed to be complete in an acrize subset of the orbital space, and the corresponding CI secular problem is solved by a direct scheme using the unitarygroup approach. With a density matrix formulation the orbital optimization step becomes independentof the size of the CI expansion_It is possible to formulate the super-C1in terms of density matrices defined only in the small active subspace; the doubly occupied orbitals (the inactivesubspace)do not enter-Further, in the unitarygroup formalism it is straightforwardand simple to obtain the necessarydensitymatrices from the symbolic formula list It then becomes possible to treat very long MC expansions,the largestso far comprising 726 configurations_ The method is demonstratedin a calculation of the potential curves for the three lowest states (IX+ 3Zc and 3fI ) of the Nr molecule, using a medium-sizedgaussianbasis set. Sevenactive orbitals were used , re: 1.108(1_098),1.309(1.287)and 1.230 yie&g be follov&g results: 0,: 8.76 (990), 2.43 (3.68) and 3.39 (4.90) eV(1113) A; 0,: 2333 (2X9), 138s (1461) and 1680(1733) cm-t. for the three states(experimentalvalues within paren-_ theses).The resultsof these calculations indicate that it is important to.consider not only the dim_ociationlimitbut also the united atom limit in partitioningthe occupied orbital space into an active and an inactive pai

1. introduction

One of the basic concepts in modem theory of electronic structure in molecular systems is the molecular orbital. Since its introduction by Hund and MttIIiken [l. 2] it has played a dominant role in qualitative as well as quantitative descriptions of what electrons are doing in molecules_ The primary tool for the calculation of molecular orbit&, in ordinary closed-shell moIecuIes, is Hartree-Fock theory [3,4], especially in the form developed for molecules by Roothaan 15-J.Such calculations are today routinely performed in many * Presentaddress: LehrstuhlErTheoretische Chemie der UniversitZtKarlsruhe,7500 Karlsruhe,Postfach 6380, Germany.

laboratories with existing ab initio self-consistent , field programs. Further, certain classes of open-shell systems can be treated with a similar technique [.5]. Hartree-Fock theory negIects the instantaneous interaction between the electrons, the so-cahed electron correlation. However, even in a more exact theory: where electron correlation is included, the molecular orbital concept survives - in.terms of the natural orbitals [6]. The natural orbital concept has become a powerful tool for analysing complex wavefunctions, and has also been used in developing methods for determining accurate wavefunc&ns. Examples are the iterative natural orbital method [TJ and the PNO-CI method ES, 9]_ , The Hartre+Fock model gtves a surpri&gly -1 accurate zeroth-order wavefunction for stabIe

-.

158

B-0. Roos et al_/CASSCF

suner-Cl

method

closed-shell molecules and also in many cases for open-shell systems. On the other handiit breaks down completely in situations where the electrons undergo substantial rearrangements, as in the formation or breaking of chemical bonds. A qualitatively correct description of these situations requires a more compIicated wavefunction, even at the zeroth-order level of approximation_ A single configuration description of the electronic structure is not sufficient here; the wavefunction must comprise several electronic configurations in order to be even qualitatively correct. This leads immediately to a mulri-configurational self-consistent field (MCSCF) theory as the natural extension of the Hartree-Fock method [lo]. The molecular orbital concept is preserved in such an extension, but the orbit& now have non-integral occupation numbers. In addition, the concept of an orbital energy can be maintained, through the use of the extended KoopmanS theorem [I l]_ MCSCF theory has been reviewed in a number of recent papers in which further details may be found jl2-14-J. Many different methods have been utilized for optimizing the orbitals in an MCSCF wavefunction. In this paper we shall deal exclusively with the so-called super-CI method developed by Grein et al. [lS, 161. The basis of this method is the Brillouin theorem [I?], in the form generalized to multi-configurational wavefunctions by Levy and Berthier [ 181. This generalized Brillouin theorem will be referred to here as the Brillouin-LevyBerthier (BLB) theorem, a name which has recently been suggested [19]_ The super-C1 method of Grein et al. uses the coefficients from a CI calculation which comprises all single excitations out of the reference MC function to construct a unitary rotation of the orbitals in each iteration_ Cheung et al. [19] have recently suggested an alternative approach, which has proved to be rapidly convergent in a number of applications [20]. In their approach, the natural orbitals (with highest occupation numbers) of the singly-excited CI (super-CI) wavefunction are chosen as rhe improved orbitals. This ensures maximum overlap between the super-C1 wavefunction and the improved MC function. -In the present work a detailed procedure will be discussed for MCSCF calculations employing a

special type of MC wavefunction for which the natural orbital super-C1 approach is a convenient choice. The MC wavefunction to be considered is constructed from two sets of occupied orbitals: the inuctiue and actice orbitals. The inactive orbitals are doubly occupied in all configurations in the MC wavefunction. Conversely, the active orbitals define a subset of the total orbital space, in which the configuration expansion is chosen to be complete.

The wavefunction

constructed in this way will be

called the complete active space (CAS) wavefunctionThis type of MC expansion, which has also been discussed by Ruedenberg and Sundberg 1211. has several conceptual as well as computational advantages. First, orbital rotations within the inactive and active subspaces leave the energy unchanged, and therefore do not need to be considered. -This leads to a simplification of the oroital optimization step, and may also, in fact, give faster convergence_ By choosing the orbitaIs to be natural orbitals these remain well-defined during the iterations, and converge to the natural orbitals of the MCSCF wavefunction. Secondly, with the use of a complete CI wavefunction in the active subspace, the sometimes difftcult choice of the dominant configurations is completely avoided and the only remaining chemical problem is the choice of active orbitals. One dilfculty may be the length of a complete CE expansion since even with a rather small number of active orbitals, the wavefunction may comprise several thousand terms: thus distributing eight electrons among eight orbitals in all possible ways yields 1764 singlet configurations if no reduction by symmetry is possible. This in turn gives rise to two computational problems. An efficient method must be found to solve the MC secular problem. The graphical unitary group approach (GUGA) developed by Shavitt [22], following the theoretical development of Paldus [23], is extremely well suited for complete CI calculations, and as it is a direct method rather long expansions can be handled in an efficient way. A GUGA program has therefore been written to solve the MC secular problem in the present work. For the more recent development of the unitary group approach, see for example, refs. [48-503.

8.0. Roos et aL/CASSCF super-Cl method The second, more stirious problem concerns the solution of the super-C1 secular problem. In the conventional super-C1 approach [15,16,19], the super-C1 wavefunction is constructed by considering first all possible single excitations out of each configuration in the MC wavefunction: obviously the number of such excitations easily becomes exceedingly large as the number of terms in the MC function increases_ Another approach is adopted here: the matrix elements between the super-C1 states are expressed directly in terms of reduced density matrices for the MC state, and the calculation thereby becomes independent of the length of the MC expansion_ In practice, this cannot bt achieved without approximating some of the matrix elements, as the exact expressions involve third-order density matrices, which would present considerable computational problems. Different levels of approximation of the super-C1 matrix elements are possible_ A practical choice is governed by the rate of convergence and the ease of computation. One possible choice to be discussed has the advantage of reducing the two-electron transformation problem substantially: as the transformation problem is one of the major bottlenecks in applications of the MCSCF method, this approximation is of special interest. In the following sections the density matrix formulated super-C1 approach will be presented and discussed in detail. The different levels of approximation will be studied and some calculations on Nz, F, and CuF, will be presented as examples.

159

tions which comprise the CAS wavefunction. In actual applications of the method these would be orbitals which have occupation numbers close to two and which are not expected to change during the chemical process studied. The active orbitals constitute the space in which the CI wavefunction is complete, that is, ail possible orbital occupations and spin-couplings consistent with the desired state are included. These orbitals have been termed reaction orbitals in the work of Ruedeiberg and Sundberg [21]: the term active is preferred here to denote a more general orbital space, which would commonly include those orbitals undergoing a radical change during the course of a chemical reaction, but could also comprise all valence orbitals (a full valence CI study) or all orbitals defining a number of excited states, etc. Labels i,j, k and I will be used for inactive orb%& t, u, u and x for active orbitals and Q, 6, c and d for secondary orbitals. The inactive, active and secondary subspaces will be designated as {Ii), (Ia> and {II), respectively_ Labels are also needed for orbitaIs beIonging to the complete space {Ii) + {Ia) -i- (II) = {I) + {II>. Letters p, q, r and s will be used in this case. The number of orbitals in each subspace is given as nn, nh and n,,, respectively, with n = nii f n,, + +_ The non-relativistic spin-independent hamiltonian for the molecular system,

(11
will be defmed in terms of the generators of the unitary group+.

2. The one-electron basis and the hamiltonian The one-electron orbital space used to build the conligurations in the CAS wavefunction wi!l be termed the primary space in the following discussion_ The remainder of the total orbital space, obtained from the atomic basis by the LCAO procedure, is termed secondary. This latter space plays no role in the final (orbital-optimized) wavefunction but is used, during the optimitation procedure, to improve the primary space. The primary space is subidivided into inactive and actit-e s&spaces_ The inactive orbitals are defined as those orbitals which are doubly occupied in all contigurawhere aL(&) is a creation (annihilation) operator for an electron in orbital yP wivlth spin quantum number K The operator Epq can also be considered as an excitation operator replacing one orbital ps with the orbital p_ and leaving the spin unchanged_ This property of Ep,, will be used later in the discussion. The-hamiltonian (1) in second quantized form is given in terms of the operators (2) as:

* For a through discussion of the unitary group and the generators of its Lie algebra see, for example, ref. [G].

160

B-0. Roos et aZ./C;ISSCF super42

method

where the one- and two-electron integrals are deiined as bps = <PI h J4>,
(P4(4 = <cp,!l) rpm1 M-1, I%#) cp,(2)>.

(4a)
WI

The generators &

obey thecommutation

relations (5)

c&* %I = &&#X - $&*-

where the sum extends over the complete configuration space 1~). Shavitt [22] has presented a detailed method for &hesolution of the secular problem which determines the expansion coefficients C,, using a direct conliguration interaction method based on the unitary group approach of Paldus [23]. In this method the matrix elements are given as

The configuration state functions (CSFs) included in the CAS wavefunction contain no orbitals from the secondary space, and orbitals from the inactive subspace will always be doubly occupied. Denoting such a CSF by I& the following relations are immediately obtained :

Gp[P>

26i,lP> (9i in {Ii))-

(6W

In the direct CI approach [24] contributions to the vector d = H C are obtained directly from the list of one- and two-electron integrals, according to
~j, = 111 c hpg C 6%; P-4 L. + 11) G (Pqb-4 P.%.S C CJ&sP (10)

From these equations and the commutation relation (5) the-following scheme is obtained for -contracting products of generators which contain secondary or inactive orbitals:

Comparing (9) and (IO) one notes that the coupIing coefficients A and B& are given as matrix ZJQ elements of single generators and the generator products by A;; = <PI Ep,, iv> 0 la)

(9; and 9j in {Ii}).

(7b)

and BS< = <~l%&~rs - &&s) Il.>(IlbJ

The relations (5)-(7) given above wil1 be used later to derive explicit expressions for BriiIouin matrix elements and for matrix elements between states which are sir&y-excited with respect to the CAS wavefunction.

3. The complete active space (CAS) CI wavefunction The wavefunction on which the present method is based is constructed as a complete CI wavefunction in the active subspace. Thus it contains all possible distributions of N, ekctrons among the I+=active orbitals, satisfying all possible spin-couplings which correspond to a total spin quantum number .S. The inactive orbitals are always doubly occupied; the number of electrons in such orbitals being N, The total number of electrons is N = N, -L-N, The CAS wavefunction is written as

In Shavitts approach a graphical representation of the Gelfand basis (the CSFs) is employed. This representation facilitates the generation of the coupling coefficients for each of the integrals in (IO), making a direct calculation of (10) possible. The information is stored as a list of Formulae, containing the coupling constants corresponding to a given integral followed by data necessary for a straightforward evaluation of the interacting CSFs Iv> and Iv>_ The only information required in performing the singly-excited state CI is the first- and second-order density matrices for the active subspace. These matrices can be generated directly from the formulae created in the CI step. The extra work involved therefore corresponds to one extra iteration in the iterative solution of the secular equation_ This can be demonstrated explicitly by starting from rhe energy expression

B-0. Roos et ol./C.GSCF suver-CI method

161

(12)
Inserting the expression (9) for the hamiltonian matrix eIements we obtain directly the first and

second-order
P.V

density matrices as

Finally, the natural orbitals are particularly suited to obtaining a chemical insight into the results of the calculation_ Ruedenberg has discussed this advantage of the natural orbitals of a CAS trpe wavefunction, and has introduced the term natural reaction ort<tals [21]_ To conform with the nomenclature introduced above they are here termed natural .actiue orbit& (NAOs).

Thus the list of coupling coefficients enables a direct calculation of these matrices; all information needed for a given matrix element is stored sequentially on the formula tape. As the CI wavefunction is compIete in the active subspace, it is invariant to unitary transformations among the active orbit&_ These orbitals are thus only defined to within such a unitary transformation. As is evident in later discussion, a useful choice is those orbitals which diagonalize the matrix D, that is, the natural orbitals. The iterative method to be discussed converges to such orbitals, which have several advantages, as follows. The calculations can be simplified by assuming this condition to be fullilled also before convergence. Normally the density matrix (or, in cases of degenerate states, a

4. The BriIlouin-Levy-Berthier theorem and the super-0 method One of the more effective methods used for the solution of the multi-configurational SCF problem is the so-called super-U method, as formulated by Grein and co-workers-[ 15,161. The method is based on the Brillouin-Levy-Berthier (BLB) or generalized Brillouin theorem [18], which can be written in second quantized form as (14) The super-C1 method begins with a calculation of the expansion coefficients in the MC wavefunction (8). A new secular problem is then constructed using IO> and the singly-excited states

symmetry-averaged

density matrix) transforms as

(SX-states), (q&;,, - G&u) IOX (IS)

the totally symmetric irreducible representation of the point group under consideration_ If it does not have this property the resuli can be said to be symmetry unstable, a situation similar to that sometimes encountered in Hartree-Fock calculations. Such instabilities are generally due to low-lying excited states; if they occur in a CASSCF calculation, this indicates that important orbitals are missing from the active subspace. However, in the case of a totally symmetric density matrix, the natural active orbitals are also symmetry orbitals. On the other hand, in a limited MCSCF calculation the orbitals often tend to localize, and thus do not transform as irreducible representations of the point group_ This reduces the extent io which symmetry can be used in the calculation. Such a problem does not arise here - the full symmetry can always be used. This is especially important for the efficient calculation of A0 integrals and the subsequent transformation of these integrals to MO basis.

as a configurational basis. The expansion coeifrcients are determined by the usual variational method. Improved orbitals can then be obtained either by performing a unitary transformation of the original MOs, based on the expansion coefficients for the SX states [IS], or by a natural orbital analysis of the resulting super-C1 wavefunction [19]_ The latter method will be used in the present work. Detailed formulae for the super-C1 density matrix elements are presented in appendix B, where they are expressed in terms of the super-C1 expansion coefficients and the fEst- and secpnd-order density matrices (13) for the CAS wavefunction. The selection of orbita!s for the next iteration from the super-C1 NOs is straightforward. The qi orbitals with the highest occupation numbers are chosen as the inactive orbitals and the next n, as the active orbitafs. The order within each group is immaterial. Again this is due to the fact that the

162

B.O. Ram et aL/CASSCF

super-CI method

CAS wavefunction is complete. None of the selection problems discussed by Cheung et al. [19] are therefore encountered in the present scheme_ In practical applications, however, the super-C1 method has some serious disadvantages. It is necessary to construct first the matrix elements between CSFs singly-excited with respect to all CSFs occuningjn (8). The number of such singlyexcited configurations can be exceedingly large if many terms are used in the multi-configurational wavefunction. The method is therefore limited to short expansions of the form (8), which in turn means that for a wavefunction of the CAS type only a few active orbitals can be used, severely limiting the applicability of the method. Further, in order to calculate these matrix elements between SX-states, a list of two-electron integrals (pqjrs) is required in which two of the indices correspond to secondary orbit&. A partial transformation of the A0 basis to such MO integrals has to be performed in each iteration; as this transformation is very time-consuming when the basis set is large it may actually become the bottleneck of such a calculation. It is, however, possible to remedy these limitations of the method to a certain extent, when the multiconfigurational wavefunction is of CAS type. As will be shown, for this case it is possible to express the BLB matrix elements directly in terms of the lirstand second-order density matrices for the CAS functions_ Approximate expressions for the interaction between the SX-states can also be obtained in this form. In this way, the need to construct explicit matrix elements between all singly-excited configurations is avoided_ Further, by introducing effective one-electron operators it is also possible to reduce the number of MO two-electron integrals to those which have at most one index in the secondary space. The BLB theorem as stated in eq. (14) corresponds to an unrestricted variation of all spinorbitals occupied in the MCSCF wavefunction IG). However, in most applications the variations are restricted to the orbitals defining the space part of the wavefunction, and the same space orbitals are used independent of the spin labeIs_ The lirst order change in energy due to unitary transformations among these orbitals is then given by (14), summed over spin. It follows that an orbital restricted BLB

theorem can be expressed in terms of the generators gPq as defined in (2), <Of fi(&,,, - &,) IO> = 0.
Expression (16) is identically zero if both the orbitals qP and ps are either inactive or active. The

(16)

action of the generators is then to produce a linear combination of states already present in IO>(or zero), and the resuIt then follows, as fi is diagonal in this space and the two generators in (16) are hermitian adjoint. This property oi(16) is a manifestation of the above observation that the wavefunction is invariant to a unitary transformation among the active orbitals. In fact, only three types of singlyexcited states interact with IO> and thus need to be included in a super-C1 calculation_ These are [i -, n> = 2-l &,j [G> (9r in {Ii) and (pnin (II)), (17a) lr -, a> = r~-r~ J?,,,]O> and Ii + r> = nr; If2 _&IO> (9i {Ii> and 9z in {Ia)), (17c) where n, = 0:: and m, = 2 - Dip_The numerical factors ensure that the SX-states, which have been constructed with the excitation operators &, are normalized_ The hermitian adjoint of these operators give identically zero when operating on IO> [cf. eqs. (6a) and (6b)] , and the LHS of eq_ (16) therefore reduces to one term. The singly-excited states defined by (17a)-(l7c) are normalized but in general not orthogonal. The overlap integrals are easily shown to be (i + czli -b> = S,a,,, (184 (W (184 (9r in [Ia] and qua in [IIj), (17b)

<t -f C+ + b> = S,,D~~/(nznJ1~, <i + tlj --. u> = 6,,(2S,, - D~~)/(mr~p,)~,

using eq_ (6) and the contraction formulae (7) Those overlap integrals not given above are identically zero. From the expressions (18) it follows that the SX-states are orthogonal if and only if the density matrix D@ for the CAS wavefunction is diagonal - that is, if the ricriue orbitals we chosen to be natural orbit&_ I& at each iteration, the natural orbitals of

B.O. Roes et al.ICASSCF

super-CI

method

163

the super-C1 wavefunction are chosen as improved orbitals, the active orbitals will be very close tonatural orbitals for the CAS wavefunction after a few iterations in the SCF procedure. Therefore, the SX-states (17) will be assumed to be orthogonal throughout the calculation, although this situation does not obtain until convergence. It should be noted, however, that this assumption will only be used during the calculation of the super-C1 expansion coefficients_ The calculation of the density matrix (cf appendix B) includes the overlap terms. It follows also from (18) that the SX-states are linearly independent except for pathological cases when the density matrix D has eigenvalues equal to zero or two. The super-C1 wavefunction, ISX>, can now be expressed in terms of the SX-states as

generator products to contain at most two Operators This result is achieved by using the commutation relation (5) together with the contractions defined by eqs. (7). When this contraction has been performed the BLB matrix elements are expressible in terms of generator products containing at most two single generators, or, equivalently, in terms of the Hurstand second-order reduced density matrices (13). The required algebraic manipulations are straightforward, although somewhat tedious, and are omitted hereThe resulting equations can be re-expmssed, in part, in terms of effective one-electron operators. Two Fock-type operators P and P (the inacrfre and active one-electron operators) will therefore be defined, whose matrix elements are given by

ISX>= Esxp>.

(1% (2W
A matrix element of the totaleffective one-electron operator P is given as the sum of these two contributions:

where the super-CL generator is defined as a linear combination of the excitation operators used in (17a)-(17c),

m-8
Here E. is the unit operator and c,, has been introduced to ensure a normalized SX wavefunction. In order to calculate the expansion coefficients c, expressions for the matrix elements between an SX-state and IO>(the BLB matrix elements) must first be obtained, as well as matrix elements between the different singly-excited states. This problem is discussed in the next section.

With the use of these operators, the BLB matrix elements corresponding to the three types of SX-states (17) are given as <Ol Ei Ii -3 n> = 21tzFse
<o/ B It 4 a> = q-l2

W-4

WW
<Ol B Ii + r> = m;l* 2F,; f c (24, Y (23c)

5. The BLB matrix elements and super-C1 expansion

coefficients
By inserting the explicit form of the hamiltonian (3) into (16), it may be noted that the BLB matrix elements can be expressed as a linear combination of matrix elements of generator products which involve at most three single generators. However, due to the deli&ions (17) of the SX-states, at least two of the six orbital indices involved will be either inactive or secondary. It is therefore possible to reduce these.

All summations in *&ese expressions are over active orbitals only. The matrix elements (23) have a very simple form, and are easily computed. The first step involves calculation of the one-electron matrices (21) and is most easily performed in a supermatrix formalism, that is, the matrices are first computed over the

164

B-0. Roes et al./CASSCF super-CI method reduce the BLB matrix elements to zero -this condition is not only necessary but also sufficient. In practice, however, a second condition also has to be fulfilled: the convergence must be sufficiently rapid to keep the number of iterations within practical limits. The above-mentioned condition can be fulfilled even if the super-CL secular problem is not solved exactly_ Several levels of approximation are possible, and an optimal choice depends on two conflicting criteria - simplicity of evaluation and rapidity of convergence. The simplest approximation which can be applied to the matrix elements between the SX-states is to assume that the third-order density matrix is given as a product of the first- and second-order density matrices, that is, to set = 0 _ L.xFIY
pl_O,_ tx,_ . ~9 tu -

A0 basis using a list of super-matrix elements, and are then transformed to the MO-basis. The computational effort corresponds exactly to the settingup of the Fock matrix in a conventional closed shell Hartree-Fock calculation [S]. In the second step, the remaining two-electron conttiburions to (23b) and (23~) are included_ This involves multiplication of two-electron integrals by the second-order density matrix Pto). Only two-electron integrals of types (azrlux) and (izrjux) are needed, and thus only one of the integral indices runs over the entire orbital space. The most time-consuming part of this step is the calculation of the second term in (23h). The calculation of all such contributions is an n,,& process, and as ,I,~ is a small number the evaluation of the BLB matrix elements will constitute only a small fraction ofthe total computational effort. For the same reason the number of second-order density matrix elements is small enough to be kept in fast storage during the whole process. However, in order to obtain an estimate of the expansion coefficients for the SX-states, it is not sufficient to calculate only the BLB matrix elements - the matrix elements between the SX-states are also needed when a super-C1 method is used. These matrix elements will in general involve products of up to four single generators, in which at least two of the eight orbital indices will correspond to inactive or secondary orbitals. Six active indices then remain after contraction, and so the matrix elements are given in terms of first-, second- and rhir&order reduced density matrices for the active space. The explicit formulae for these matrix elements have been presented earlier [25]. For completeness they are, however, given again in appendix .4. Obtaining third-order reduced density matrices from a Shavitt grdph is, however, a complicated procedure, and a simplification of this part of the c&uIation is therefore highly desirable. It may be noted that the iterative procedure used in an MCSCF calculation must obey one obvious condition:% should converge, that is, it should
* .4ctuaIIy only a component ofthe third order density

This will affect only a small number of matrix elements (cE appendix A). All matrix elements are then expressed in terms of first- and second-order density matrices. However, they still contain twoelectron integrals with two orbital indices outside the active subspace. The next level of approximation, which has been investigated in detail, is to assume that the firstorder density matrix is diagonal D ,Y = 0 I6 Ill1

(25)

and to approximate the second-order density matrix by the occupation numbers n,:

PO) ,YYI = --qtlpu

(t # u),

(26W

matrix is needed, according to the definition given in appendix At but for the sake of simplicity the matrix Q
defined in eq. {A-l) xvi11 in the text be referred to as the

third order density matrix_

all other matrix elements being set to zero. Thus onIy the Coulomb and exchange parts of the secondorder density matrix are kept. The matrix elements between the SX-states now take an especiaIly simpIe form. They are given in eqs. (A.4aHA.4f) in appendix A. These matrix eIements are a11expressed in terms of the Fock operator (22) plus a few additional two-electron integrals_ This level of approximation wili be denoted Al_ An obvious third level of approximation (denoted A2) is then to neglect these integrals and keep only the Fock matrix contributions. As wil1 be ihustrated in the next section this

B.0. Roes et aLfCASSCF super-CI method

165

approximation has, in fact, performed quite satisfactorily in a number of applications. It has the great advantage that the only transformed twoelectron integrals necessary are those used to construct the BLB matrix elements (23), and they contain only one orbital index outside the small active subspace. In the next section some examples will be given to illustrate the convergence properties of the approximate super-C1 approach. It is clear from the timing data given that the choice between the different levels of approximation depend to a large extent on how the rate of convergence changes compared to the reduction in transformation time. Another approach which could be used in cases where the more approximate scheme does not converge satisfactorily is to use the same interaction matrix for several iterations and only change the BLB matrix elements.

6. Computational results The first test applications of the CASSCF super-C1 method have been directed towards examining the convergence behaviour of the method rather than the use of a very large number of configurations in IO>. In this section, timing aspects of several calculations will be discussed; a fuller treatment of results obtained for several Na potential curves is given in the next section. Table 1 Iists the convergence of CASSCF calculations on the ground state of N, at R = 2.075 au. Results are presented for both the AL and A2 schemes described above. The atomic basis set used consisted of van Duijneveldts (9s, 5p) basis [263, contracted as (5,1,1,1,1/3,1.1), and augmented with one d function, exponent 095. Orbitals from a closed-shell SCF calculation which gave au, were used as a starting guess E SCF= - IO.%96898 for both CASSCF calculations. The active space comprised the 2q,, 3rr,, In,, lxr and 3~rorbit&, giving a CASSCF wavefunction with 80 configurations_ It is quite evident from table 1 that the convergence behavlour of the A2 scheme is in no way inferior to that of Al. Further the A2 scheme is more than three times faster (CLtable 2): as expected the major part of this difference results from the

need to transform integrals in which two indices span the iull orbital space, and which results in a four-fold increase in CPU-time during the transformation step. For the A2 case this step requires around 60 o? of the time used in one iteration, while for Al almost 85 % of the iteration time is consumed by the transformation. In both cases, the single excitation CI step requires a rather small fraction of the time. The difference between the Al and A2 timings rellects mainly the different number of diagonalization iterations required during this step: it appears that the somewhat simpler form of the matrix elements used in the A2 scheme results in a matrix which is more diagonally dominant than is the case for Al_ Owing to the appearance, during the iterations, of some diagonal super-C1 matrix elements wirh values below that of IO>, both schemes require the introduction of a level shift. A value of 2.0 au was used for the N, calculations; although this is rather larger than the values commonly used in SCF calculations i27], our experience has been that Table 1 Convergenceof the CASSCF super-Cl method (N2 with 40 basis function, R = 2.075 aur Iteration E 1 max(@) E max (c)b Al A2

- 107.07967 0.312 - 108.88821 3.584 109.03442 109.05770 109.06775 109.09028 109.10552 109.11085 109-l1228 0.288 2481 0.744 0.379 0.176 0.148 0.066

- 107.07967 0.316 - 108.88637 3.084

109-03399 0.302 109_05753 3.488 109.06775 0.695 109_08604 0.319 109.10238 0250 109.11038 0.172

3 4 : 7 8 9

10
11 12 13 14 15

- 109.11258 0.044

109.11199 O-056 - 109.11245 0.029

109.11260 MIS 109.11266 OIlO6 109.11269 0.003 109.11271 O_OOl 109.11272 <lo-

- 109.11265 O-032 - 109.1126b 0.022 - 109_11270 Ml8 - 109_11271 0.012 -109.11271 0.011

Starting vectors from an SCF calculation,on the closed shell ground state was used. bJMaximum super-C1 coelficient. c, Better couvergcucxz than 10m5 is not meaningful in a

single precision calculation on a UNIVAC 1108 computer which has a word length of 36 bits.

166

B.G. Roos et al_fCASSCF super43

method

Table 2 Timing data for some CASSCF super-C1 calculations Molecule Number
Of aciive

orbitals
N, N, N1 % F, CuF, 7 7 6 9 7 12

electrons
S Is 6 8 6 21

of ac:ive

Namber

Number rations
so 80 32 726 80 86

of contigu-

of approximation Al A2 A2 A2 A2 A2

Level

CPU time per iteration

transformation 21 5 4 6 5 15 CI 2 2 1 2fd 2 11 super-C1 2 I 1 1 1 3

Number of iterations
IS I5 17 16 20

Basis ,,I size

40 40 40 40 40

50

starting vectors were obtained from SCF calculations. Cl Using orb&Is from the seven active orbit& calculation. First iteration. Decreases with the number of iterations (see text).

In seconds on a UNIVAC I lOO/SO computer_ bf Number of iterations needed for convergence to IO- au in energy and IO- in the super-C1 coeflicients. In all cases

smaller values may not be adequate, and in any case seldom provide much improvement in convergence_ As a final note on the timing for the super-C1 step, it is worth reiterating that this step is expected to be quite rapid in general, as it involves obtaining only the lowest root from a small (dimension nlnlr f nun,) CI problem_ Some additional calculatioris are also given, as a further illustration of timing, in table 2. The first calculation is again the ground state of N,. but with the 2c,, orbital transferred from the active to the inactive subspace (see below). This gives six active orbitals and six active electrons, and is the smallest possible active subspace which properly describes the dissociation of N, into two nitrogen atoms. The dimension of the CI problem is now reduced from 80 to 32 (D,, symmetry was used in all calculations presented in tabie 2). The timing for the CI step is correspondingly decreased, whiIe the other times remain almost unchanged; the transformation time is a little smaller owing to the decrease in active subspace. A preliminary calculation has also been perperformed on N,(X XL j with eight electrons distributed among nine active orbitals; in addirion to the calculation using seven active orbitals, the 2% orbitals are also included in the active subspace. The CASCI wavefkction now comprises 726 configuration state functions. The timing is dominated by the CI step. which in the first iteration needs

21 s. In the direct CI scheme this time is proportional to the number of iterations necessary to converge on the CI energy and wavefunction. Six such CI iterations were needed in the first SCF iterations_ Near SCF convergence fewer CI iterations are needed since the calculation can start from the CI coefficients of the preceding SCF iteration_ Thus, in the last iteration only two CI iterations were needed and the CPU time used in the CI section dropped to 7 s. Most noticeable in this calculation is that the time used in the orbital optimization step has not changed compared to the smaller N, calculations. This part of the calculation is indeed independent of the size of the CI wavefunction and varies only slowly with the size of the active subspace. This is a unique feature of the density matrix formulation of the super-C1 approach as it has been presented in

this work.
The two other results in table 2 represent calculation on the F, molecule, and, as a somewhat Iarger test problem, the CuF, molecule_ The F, calculationwas performed as part of an attempt to improve on previous, unsatisfactory MCSCF and CI calculations [28] which were felt to be in error due to the omission of configurations involving the 2x,, orbital from the CI reference space. Accordingly,;

an active subspace of seven orbitals, 3~ 3crU,, tzg, 4mgand 2% was used. With six active electrons, this defines a similar CI problem to the larger N,

B-0. Roes et ol.lCASSCF

super-CZ method

167

calculation. An atomic basis analogous to that used for Ne was used. Timings are generally similar to the Nz calculation; convergence was obtained in 16 iterations from a closed-shell SCF starting guess, despite the fact that the latter is a rather poor approximation for Fz. Copper halides, such as CuF,, and their electronic spectra form an area of current interest in this laboratory [29]. For the CASSCF calculations on the ground (C,) state, Wachters atomic basis set [30] was used for Cu, contracted to (SJ,l,l,l,l/ 6,1,1,1/4,1) and the Dunning basis [31] contracted to (6,1,1,1/4,1) for F, giving 50 contracted functions. Molecular orbit& arising from the Cu 3d and 4s orbitals and the F 2p orbitaIs (in MO notation: 6crg,,709, 8crg, 3uU,,1&4a,, 5a,, lg and 2nJ were included in the active subspace, giving 21 electrons in 12 orbitals. Timing data for the calculation are given in table 2 Even here, the transformation step needed for the A2 scheme comprises only half of the total time -whilethe timing contribution for the super-C1 step is essentially negIigible_Convergence was achieved in 20 iterations, using orbitals from a closed-shell SCF calculation on CuF; as a starting guess. This guess is a particularly poor one as the partitioning into occupied and virtual orbitals in CuF; is rather different to that of CuF,, which does not obey the au&au principle at the RHF level. Thus the genera1 conclusions from the timing data discussed above provide strong support for the contention of earlier sections, that in calculations using large active subspaces and long CI expansions, the rate-determining step will be the CI calculation itself, or, for large atomic basis sets, the transformation step. while the super-C1 step should require only a small fraction of the total time. The rate of convergence achieved in the calculations presented in table 2 is acceptable, but not comparable to that which can be obtained in a full second-order process (see e-g. ref. [32]). However, at least in the Nz case in table 2, a second-order process would have to converge in less than six iterations to be competitive with the A2 scheme in timing, since the iteration time increases with a factor of three. The calculations of the potential curves for N,, to be discussed in the next section, show even better convergence behaviour, since in this case starting vectors from nearby points on

the curves could be used. Generally, these calculations converge in about 10 iterations. 7. Potential curves for N1 As an example of the capabilities of the CASSCF method, potential curves for the three lowest lying states (X1x:, A3Z:, B3H,) of the N, molecule have been calculated. Although MCSCF calculations on excited states have appeared [33], there has been surprisingly little work published on the lower states (although, see ref. [34] for details of several unpublished studies). The three states listed provide a variety of conditions under which the CASSCF method can be tested, as the ground state is dominantly a closed shell near equilibrium, the A3x,f state is open-shell, and the B3fl, state requires density matrix averaging if full axiai symmetry is to be retained. Further, while the X and A states dissociate into two ground (SO)state nitrogen atoms, the B state dissociates into the ?S + Do channel [35], and hence calculation of these curves also provides a test of the ability of the CASSCF method in reproducing the splitting between different asymptotic limits. An obvious choice of active subspace for such calculations would comprise the 3cg,, lx,, lx, and 3cU,orbitals, denoted CASSCF-6; such a choice may equivalently be regarded as including the atom 2p shells in the active subspace, and automatically includes all conligurations required for dissociation into the proper atomic states. Further extension of this active subspace is discussed below. The atomic basis set used was that given in the previous section. The A2 super-CT scheme was used for all calculations. The number of configurations included in [O> was 32,27 and 24 for the three states, respectively. (D,, symmetry was used in the calculations instead of D,, for computational reasons_) A grid size of O-05 au was used around the energy minimum region for each of the states. The resulting potential curves are plotted in fig_ 1, and spectroscopic constants obtained from these curves by means of a Dunham analysis [36], together with experimental values [34] are given in table 3_ A fifth-order polynomial, fitted to seven points around the minimum, was used in this analysis_

168
e-J t

B.O. Roos et nL/CASSCF

super-CI

method

:I
7

0 _I
-1

1 I
i

-2

-3

-4

-5

-6

-7

i J i
I
i I

1 -i 1 i -i

j
-e -4

35

Fi_e I. Calculated and experimentai potential curves for the three lowest states ofthe Nz molecule: (---) give the CASSCF-6 resuItsr (----) the CASSCF-7 results and (.__) the ucpelimental curves [347. Tabk 3 Spectroscopic constants for NIXI
XZ; n, = 6 n, = 7 exp

For the ground state, the basis set used gives a R, value of 1.069 A at the SCF level [37], and so there is a substantial improvement over the SCF value at the CASSCF level. The dissociation energy of the ground state, a notorious failure of the SCF method, which gives an error of 4.5 eV, or almost 50 T<of D, [38], is also brought into more satisfactory agreement with experiment using the CASSCF method_ For comparison, Lie and Clementi, in an MCSCF calculation using only the the configurations needed for proper dissociation, obtained a D, value of 727 eV 1391 in spite of the fact that the calculations were performed with a much larger basis set (7s,4p and 2d)_ The dissociation energy results for the excited states are less satisfactory_ Although the absolute error for the A state D, is approximately the same as that for the ground state, the relative error is disappointingIy large. The entire curve for the B%, state is in rather poor agreement with experiment, displaying a maximum at 2.0 A, and a minimum which is much to shallow_ The asymptotic limit, however, is in fair agreement with experiment, and is actually very close to the HF limit of 2.85 eV (no correiation is included in the asymptotic limit wavefunction). A rationalization of some of these difficulties, and their resoktion, is presedted in the following discussion, and involves enlargement of the active subspace.

A3P
Y

B n, = 7 1.309 2.432 1385.0 15-15 1.405 eXp 1.287 3.680 1460.6 13.87 I_455

%I,
n, = 7 1.230 3.386 1679.6 14.39 1.592 0.0179 5.72 expb) 1213 4.896 1733.4 14.12 1.637 0.0179 s_9d9d

n, = 6 1.307 2.268 1410.1 19.59 lAlO

n, = 6 IL243 2.171 1543.0 17-75 1557

R,

(A)

I.105
8.574 2366.3 16.93 l-970 5.46

1.108
8.759 2332.6 14.55 1962 555

1.098
9905 2358.6 14.32 199s

D, (ev) w= (cm-l) -cr, (cm - f B, En-) q (cm-) d,(cm- x IOe6)

0.0164

0.0175

0.0173
5_76d

0.0203
5.64

o_oi95
5-78

0.0180
6_ld

0.0224
6.42

= First and second columns give calculated results with six and seven active orbitals, respectively. bJRef. [34]. Centrifugal distortion constant_ d Ret [Sl]_

B.0. Roos et al./CASSCF

super-CI method

169

Most discussion of MCSCF configuration selection (or, in this context, the choice of the active subspace) has concentrated on the need to describe the separated atom channels properly (40,413. However, recent experience in calculations on F, and Li, [28] has suggested that consideration of the united atom channel can be an important reqniremenr in the choice of active subspace, and a detailed discussion of this problem will be given elsewhere [42]. For the present case, the X, A and B states correlate adiabatically with the lowest D, P and 3P states of Si [34,43]. With the above choice of a six-orbital active subspace, it is straightforward to determine from the correlation diagram [34] that the ground state actually correlates with a mixture of Si rD (ML. = 0) and S, while the A and B states correlate with a complicated mixture of excited states involving the Si 3d orb&&_ One reason for these problems is the inclusion of the 2rrUMO in the inactive subspace, as this orbital tends to Si 3p,, which is actually higher in energy than the united atom limits of 3ba (Si 3s) and In, (Si 3~). The Si 3p= orbital has an occupation number of less than two in all three united atom limits, which suggests that the 2~~ orbital should be transferred from the inactive to the active subspace. CASSCF calculations have been performed with an active subspace of seven orbitals, obtained by augmenting the list given above with 2q, (CASSCF-7). The potential curves obtained are also plotted in fig. 1 and the spectroscopic constants are given in table 3. For all three states the inclusion of 2~ into the active subspace leads to improved resuhs. The improvement in the liI, resuhs is, however. particularly noteworthy_ The erior in the R, value decreases from 0.030 A to 0.017 %L and the binding energy (0.) which in the CASSCF-6 calculation was only 44% of the experimental value is now 68%. Substantial improvements are also obtained for the spectroscopic constants as is evident from the last columns of table 3. That the 2uU orbital is especially important in the B state can also be inferred from the occupation numbers. The 2c_ orbital has an occupation number of 1.99 at R, for the X and A states. At the equilibrium bond distance for the B state this number is 195; 0.05 electrons have been transferred to the 35 orbital.

However, even if the active 20, orbital leads to a considerable improvement in the results for the B state, some of the spectroscopic constants are still far from the experimental values. The error in the D, value is, for example, 1.15 eV for the X state. Even ifthis can be attributed to a large extent to the lack of dynamical correlation in the present cafculations, it is not unlikely that the 2% orbital also plays an important role in the electronic structure of the Na molecule. This orbital was found to be of particular importance in recent studies on F, [28]_ Calculations of a number of Nz potential curves, where 2q is included into the active subspace are presently under way. These calculations employ a more extended basis set than that used here+, The vertical excitation energies, XZ: + A3Zi and XrE: + B31T, can he obtained from the potential curves. The calculated values are 6.33 and 899 eV, to be compared with the experimental values 6.2 and 8.1 eV, respectively 134]_ The larger error for the X -, B excitation is partly explained by the error in the asymptotic energy differences The D-% energy difference obtained here is 2.88 eV, which is close to the HF value 2.85 eV [44]. The experimental splitting is only 239 eV [34], and the difference is due to dynamical correlation effects which are not iuchtded in the present treatment I& however, the 3Hs potential curve is shifted to the correct asymptotic limit the vertical X -, B energy difference is lowered to X.50 eV, in much closer ageement with experiment_ In addition, the CASSCF-7 calcuIations result in a potential maximum for the B state at about 20 A (a small maximum is actuaiIy also obtained for the A state)_ The appearance of such maxima has been noted before for Or [41] and CH c45], they are caused by a breakdown of spherical symmetry in the separated atoms - the calculations at the asymptotic limit are performed in linear rather than spherical symmetry. This results in an artificial lowering of the atomic energies caused by a contamination of d, into s, and different radial dependencies in 2p, and 2p,. This is especially serious for atomic states rvitlt

* A preliminary calculation on the X state using9 active

orbitals (including2zJ and the same basis set as before givesa 0. value of 9-05 eV.

170

B-0. Roes et al./CASSCF

super-CI method

L quantum number different from zero, where the different M, and Ms components are found to be non-degenerate [41]. There does not seem to be any simple way of avoiding this at the MCSCF Ievel, even though the inclusion of singly-excited conligurations (first-order conliguration interaction) seems to reduce the splitting to very small values, and removes the maxima from the potential curves [41]. This problem has, however, to be borne in mind when the MCSCF method is used to caIcu!ate energy surfaces for chemical reactions, for which either the reactants or the products possess higher symmetry than the compound system.

8. Concluding remarks The density matrix formulation of the super-C1 MCSCF method presented in this report, reveals some of the difficulties inherent in this method. The matrix elements between the SX-states involve the third order density man-lx of the reference function and therefore camtot be evaluated without encountering computational difliculties. On the other hand, the conventional technique where all singly-excited states from each of the configurations comprising the reference function are considered explicitly severely limits the number of such configurations. In this work an attempt has been made to resolve this problem by approximating the matrix elementsIt has been shown that an approximation is possible which also, to a large extenl removes one of the other bottlenecks of MCSCF calculations, the four index transformation of molecular integrals_ The price to be paid for this simplification is a decreased convergence rate. In the examples given here, however, the convergence is satisfactory and it is doubtful whether a full second-order process would be competitive_ This may not always be the case, of course_ The super-C1 technique is naturally not the only orbital optimization method which can be used in the CASSCF method. In a forthcoming paper it will be shown how it can be apphed using a fu!l secondorder process: the Newton-Raphson method [25,46]. The most important result obtained from the present formulation of the MCSCF method is

probably the negligible amount of time needed for the orbital optimization step and the application of the graphical unitary group approach to solution of the CI problem and the construction of the density matrices_ As a result it is now possible to treat long MC expansions in a computationaly efficient way. Another important aspect is the use of a complete: expansion in an active orbital subspace This offers certain computational advantages, in that the energy is unaffected by unitary mixing of these orbitals. Further, this form of expansion has additional benefits when it is desired to treat states which are not the lowest of their symmetry type. Thus, assuming that the active space is chosen so that states lower in energy than the one in question are properly described, not only is the CASCi wavefunction orthogonal to the lower roots, but the SX wavefunction for the state in question is also orthogonal to these lower roots. Hence modification of the wavefunction using the SX natural orbitals should not improve any lower state at the expense of the desired state-in principle, no constraining of the orbital variations need be imposed. This feature is unique to the case for which the MC expansion is complete in the active space. Finally, the benefits of such a complete expansion are not only computational_ As in the SCF method the orbital concept again occupies a central position and it is possible to discuss the chemical problem entirely in terms of the molecular orbitals and their occupation numbers. Obviously it is not possible to include dynamical correlation effects into the CASSCF wavefunction, since this would require a very large active subspace. The CASSCF wavefunction is, however, an excellent zeroth-order ipproximation, and it can be expected that the remaining correlation effects will vary smoothly with the parameters which determine an energy hypersurface. An approximate treatment of these effects might therefore be sufficient. It is, for example, straightforward to include, to second order, the correlation energy arising from all primary to secondary double replacements into the CASSCF energy. A large fraction of the remaining correlation energy can probably be recovered by such a calculation

-B-O_ Raos et nL/CASSCF

Super-U merho< -.

171

Acknowledgement One of us (P.R.T.) wishes to actiowledge the support of a Commonwealth Scientific and Industrial Research Organization Postdoctoral Appendix A: Expressions for the SX-state matrix elements

Fellowshib during part of this work. This work was supported by a grant from the Swedish Natural Science Research council.

The matrix elements between the SX-states (17) given below are expressed in terms of the Fock operators delined in expressions (21) and (22). the density matrices over the active space, as defined in (13), and the two-electron integrals (4b). In addition a component of the third order density matrix, according to QrWzru = t <Ol(L% - &.&&,. 1% (A-1)

is needed_ The matrix elements are obtained using the definition (3) of the hamiltonian and the ^contractionW formulae (7) for the generators. Defining B=R-<OlfilO>
we now obtain,

(A-2)

+ 26&

- (26,, - D,JF$

- &, G

- &.) [(utjij) - 2(ujlti)] - c(&lr - D,d [(uulfj~- 2(uilrqJ] c

(A.3a) <i 4 tl ii b + U> (2Jl1,) * = 26,F,, - Z(atlii) - Z(ajlti) - 26, <i -b tl fi Iu -+ Q> (JJGJJzJ
=

- 5, c D&,

+ g Q,[(au lij) - S(ujlui)] (A.3b)

C P,,,(oulLy), L-.X_
-

CLllatld

- 2k7uIti)l -

2& PtcwXaxlui),

(A.3c) (A.3d)

<i--f aI I? b + 6) = 6uF,,b - &.F,j + Z(ailbj) - (ublij), <i + UIB If + b) (24)

<r + aI & Iu + b> (n$y)=

= c D,,[2(4bu)
c = D,,F:,

- (abliu)] - S,, CD&,

c D,,D,.x]F;x

- 24,

r-.x7

C ~,,,(uilxy),

(k3e)

+ S,, c [2P,,, L;X

(A-3f) These matrix elements can be approximated by assuming the first-order density matrix to be diagonal, and by using an occupation number approximation of the second- and third-order density matrices according to (26). The following simplified formulae are then obtained_ (i tl ET b -+ u> (JJJpzp = a,-[+, wdl &3 + JJvL]

F,, - 6&Fij

+ &m.[2(ti~u~~
llu(tUIUU)],

(ru lo)]

- qj 2

tzPGnr[(ttlij)

(rilQl]

Satj[tVL(l

&I)

m*bw] [&(ft[tN)+

(A.42)

172

B-0. Roos et al./CASSCF

super-CTmethod

.. . - (A_4b)
+w (A-44

<i+tlA[i+a>(2inJ
<i-+t~A~z4-+a>(!72,?&)

u2 = b,fn,F,, + fn*nz(at~ tr) + ?q[2(Qjlfi) - (&,r)],

I+ = fn,m,[Z(au[ ti) (uf [ui) + &(at Iti)], (abl$, &,F,J + &,m,(trlrO,

<i + uI fi b + 6) = bijFnb -

&,Fij + Z(ailbj) -

<i + aI B It + b) (2n*)* = ?,[2(ailbt)

(ublri) -

O-3

In all these expressions n, = D, and m, = 2 - DrI, and the index (0) has been omitted from the density matrix elements in order to simphfy the notation_

Appendix B: The super-C1 density matrix elements The elements of the super-C1 density matrix are defined as
DE = <O[ J&$& [O> (B-1)

with the super-C1 generator Enr defined in (20) In order to simplify the notations the following unnormalized super-C1 coefiicients are used:
a, = co, a, = cp, - 1:2 ,

a, = ci,2-:,

a,= = c&:.

(B-2)

The super-C1 density matrix elemects are obtained using the same technique as was used for the SX state matrix elements_ Expressed in terms of the coefficients (B2) and the first- and second-order density matrices (D and P) for the CAS wavefunction we obtain: (B.3a) (B.3b) (B.3c)

ez

= 2 C ai,uib i- c a,,a,,D,,.

I.

References
El] [2] [3] [4] [s] E Hund, A. Physik 51 (1928) 759. RS. Mulliken, Phys. Rev_ 32 (1928) 168. D.R_ Hat-tree. Cambridge PhiI_ Sot. 24 (1926) 89. V. Fock, 2. Physik 61 (1930) 126. CCJ. Roothaan, Rev. Mod. Phys. 23 (1951) 69; 32 (1960) 179. [6] P-0. LBwdin, Phys. Rev. 97 (1955) 1474. [7] C-F_ Bender and E_R_ Davidson_ Phys Rev_ 139 (1969) 23. R Ahhichs. H. Lischka V_ Staemmler and W. KUQehigg, J. Chem. Phys. 62 (1975) 1225. [lo] D.R Hat-tree, W. Hartree and B. Swirlas, Phil. Trans. R Sot. London Ser. A238 (1939) 299. [ll] D-W. Smith and 0-W. Day, J. Chem. Phys. 62 (1975) 113. [ 123 J_ Hinze, J. Chem Phys. 59 (1973) 6424. [S] [9]

\V_ Meyer, Intern J_ Quantum Chem. SS (1971) 341.

[I33

AC_ Wahl and G_ Das in: Methods of electronic

B.O. Roos et al.ICASSCF structure theory, ed. H.F. Schaefer III (Plenum Press, New York, 1977). r143 E_ Dalgaard and P. J~rgenses J. Chem. Phys. 69 (1978) 3833. 1151 F_ Grein and T-C_ Chang. Chem. Phys. Letters 12 (1971) 44. cw F_ Grein and A_ Banejee, Intern_ L Quantum ChemS9 (1975) 147; J. Chem. Phys. 66 (1977) 10.54. L_ Brillouin, Act Sci Ind. 71(1933) 159. !371 fP1 B. Levy and G. Berth&. Intern J. Quantum Chem 2 (1968) 307. Cl91L.M. Cheung, ST. Elbert and K Rucdenberg, Intern. J. Quantum Chem 14 (1979) 1069. PO3ST Elberf private communication_ PII K_ Ruedenberg and K-R. Sundberg, in: Quantum science, methods and structure, eds. J.-C. Calais, 0. Goscinski, J. Linderberg and Y_ ijhm (Plenum Press, New York, 1976). ca I. Shavitt, Intern. J. Quantum Chem Sll(1977) 131; s12 (1978) 5. PI I Paldus, in: Theoretical chemistry: advances and perspectives, eds. H. Eyring and D_ Henderson, Vol. 2 (Academic Press, New York, 1976) p. 131. c241 B. Roes, Chem. Phys. Letters 15 (1972) 153; B. Roos and P. Siegbahn, in: Methods ofekctronic structure theory, ed. H-F. Schaefer III (Plenum Press, New York, 1977). P. Siegbahn, A. Heibcrg, B. Roos and B. Levy, 1251 Phys. Ser. 21 (1980) 323. PI F.B. Van Duijneveldt, IBM Research Report RJ945 (1971). c271 M.F. Guest and V.R. Saunders, Mol. Phys. 28 (1974) 8:9. WI B. JBnsson, B. Roes, P. Siegbahn and P. Taylor, to be published_ PI P.C. de Mello, M. Hehcnberger, S_ Larsson and M. Zemer, Inorg. Chem, to be published.

super-CI

method ._.

173

[30] A.J.H. Wachter, J. Chem. Phys. 52 (1970)~1033. [31] T.H. Dunning, J. Chem. Phys. 53 (1970) 2823. [32] D-L.. Yeager and P. Jergensen, J_ Chem. Phys 71 (1979)755. [33] M. Krauss and D.B. Neumann, Mol. Phys. 32 (1976) iO1. [34] A_ Lofthus and P-H_ Krupenie, J_ Phys. Chem. ReT. Data 6 (1977) 113. [35] G. Heraberg, Spectra of diatomic molecules (Van Nostrand, Princeton, 1950). [36] J-L- Dunham. Phys. Rev_ 41(1932) 721. [37] R. Ahhichs, H. Lischka, B. Zurawski and W. Kutxelnigg, J_ Chem Phys. 63 (1975) 4685 [38] W. Butscher, S.-K Shih. RJ. Buenker and SD. Peyerimhoff, Chem. Phys. Letters 52 (1978) 457. [39] G.C. Lie and E Ciementi, J. Chem. Phys. 60 (1974) 1288. [49] F.P. Billingslcy II and M. Kmuss, J. Chea Phyr 60 (1974) 4130. [41] R.P. Saxon and B. Liu, J. Chem. Phys. 67 (1977) 5432 1421 P.R. Taylor, to be published. [43] AC. Hurley, Introduction to the electron theory of small molecules (Academic Press, New York, 1976). [44] E Clementi and C. Roetti, At. Data Nucl. Data Tables 14 (1974) 177. [45] G.C. Lie, J. Hiuze and B. Liu, J. Chem. Phys. 59 (1973) 1872 [46] P. Siegbahn, A Heiberg and B. Roes, to he published [47] M_ Moshinsky, Group theory and the many-body problem (Gordon and Breach, New York, 1968). [48] I. Shavitt, Technical Report (Batelle Columbus Laboratories, 1979). [49] B-R. Brooks and H.F. Schaefer III, J. Chem. Phys. 70 (1979) 5092. [SO] P-EM_ Siegbabn, J. Chem. Phys. 72 (1980) 1647. [Sl] K.P. Huber and G. Henberg, Constants ofdiatomic molecules (Van Nostrand, Princeton, 1979).